The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains explicit defensive instructions in references/content-to-skill.md and references/content-to-skill-corpus-design.md directing the agent to treat processed transcript text as data only, ensuring that any embedded commands or override attempts within the source material do not alter the agent's core instructions.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process external content libraries, which is a recognized ingestion point for untrusted data. The risk is mitigated through a structured approach documented across several files:
Ingestion points: Specifically defined in references/content-to-skill.md as owned content libraries (transcripts, podcasts, webinars).
Boundary markers: Instructions require that source material be treated as data and clearly distinguished from editorial synthesis.
Capability inventory: The skill is entirely instructional and does not request or invoke tools with network or filesystem-writing capabilities.
Sanitization: The workflow mandates manual review, pilot evaluation, and verification of extracted claims against their original source.
[DATA_EXFILTRATION]: The skill actively encourages data privacy and safe handling. references/content-to-skill-release.md provides explicit guidance to exclude local caches, hidden credentials, and private customer examples from any distributed skill artifacts.
[SAFE]: Adversarial prompt strings found in evals/evals.json are test cases used to verify the robustness of the agent's safety guardrails during development and do not represent malicious intent within the skill code.

lead-magnets